ARTICLEdoi:10.1038/nature11981

Dynamic regulatory network controllingTH17 cell differentiationNir Yosef1,2*, Alex K. Shalek3*, Jellert T. Gaublomme3*, Hulin Jin2, Youjin Lee2, Amit Awasthi2{, Chuan Wu2, Katarzyna Karwacz2,Sheng Xiao2, Marsela Jorgolli3, David Gennert1, Rahul Satija1, Arvind Shakya4, Diana Y. Lu1, John J. Trombetta1, Meenu R. Pillai5,Peter J. Ratcliffe6, Mathew L. Coleman6, Mark Bix5, Dean Tantin4, Hongkun Park1,3, Vijay K. Kuchroo1,2 & Aviv Regev1,7

Despite their importance, the molecular circuits that control the differentiation of naive T cells remain largely unknown.Recent studies that reconstructed regulatory networks in mammalian cells have focused on short-term responses andrelied on perturbation-based approaches that cannot be readily applied to primary T cells. Here we combinetranscriptional profiling at high temporal resolution, novel computational algorithms, and innovative nanowire-based perturbation tools to systematically derive and experimentally validate a model of the dynamic regulatorynetwork that controls the differentiation of mouse TH17 cells, a proinflammatory T-cell subset that has beenimplicated in the pathogenesis of multiple autoimmune diseases. The TH17 transcriptional network consists of twoself-reinforcing, but mutually antagonistic, modules, with 12 novel regulators, the coupled action of which may beessential for maintaining the balance between TH17 and other CD41 T-cell subsets. Our study identifies and validates 39regulatory factors, embeds them within a comprehensive temporal network and reveals its organizational principles; italso highlights novel drug targets for controlling TH17 cell differentiation.

Effective coordination of the immune system requires careful bal-ancing of distinct pro-inflammatory and regulatory CD41 helperT-cell populations. Among those, pro-inflammatory IL-17 producingTH17 cells have a key role in the defence against extracellular patho-gens and have also been implicated in the induction of several auto-immune diseases1. TH17 differentiation from naive T cells can betriggered in vitro by the cytokines TGF-b1 and IL-6. Whereas TGF-b1 alone induces Foxp31 regulatory T cells (Treg cells)2, the presenceof IL-6 inhibits Treg development and induces TH17 differentiation1.

Much remains unknown about the regulatory network that controlsthe differentiation of TH17 cells3,4. Developmentally, as TGF-b isrequired for both TH17 and induced Treg differentiation, it is not fullyunderstood how balance is achieved between them or how IL-6 pro-duces a bias towards TH17 differentiation1. Functionally, it is unclearhow the pro-inflammatory status of TH17 cells is held in check by theimmunosuppressive cytokine IL-10 (refs 3, 4). Finally, many of the keyregulators and interactions that drive the development of TH17 cellsremain unknown5.

Recent studies have demonstrated the power of coupling systematicprofiling with perturbation for deciphering mammalian regulatorycircuits6–9. Most of these studies have relied upon computationalcircuit-reconstruction algorithms that assume one ‘fixed’ network.TH17 differentiation, however, spans several days, during which thecomponents and wiring of the regulatory network probably change.Furthermore, naive T cells and TH17 cells cannot be transfected effec-tively in vitro by traditional methods without changing their pheno-type or function, thus limiting the effectiveness of perturbationstrategies for inhibiting gene expression.

Here we address these limitations by combining transcriptionalprofiling, novel computational methods and nanowire-based short

interfering RNA (siRNA) delivery10 (Fig. 1a) to construct and validatethe transcriptional network of TH17 differentiation. The reconstruc-ted model is organized into two coupled, antagonistic and denselyintra-connected modules, one promoting and the other suppressingthe TH17 program. The model highlights 12 novel regulators, thefunction of which we further characterized by their effects on globalgene expression, DNA binding profiles, or TH17 differentiation inknockout mice.

A transcriptional time course of TH17 differentiationWe induced the differentiation of naive CD41 T cells into TH17 cellsusing TGF-b1 and IL-6, and measured transcriptional profiles usingmicroarrays at 18 time points along a 72-h time course (Fig. 1,Supplementary Fig. 1a–c and Methods). As controls, we measuredmRNA profiles for cells that were activated without the addition ofdifferentiating cytokines (TH0). We identified 1,291 genes that weredifferentially expressed specifically during TH17 differentiation(Methods and Supplementary Table 1) and partitioned them into20 co-expression clusters (k-means clustering; Methods, Fig. 1b andSupplementary Fig. 2) with distinct temporal profiles. We used theseclusters to characterize the response and reconstruct a regulatorynetwork model, as described below (Fig. 2a).

Three main waves of transcription and differentiationThere are three transcriptional phases as the cells transition froma naive-like state (t 5 0.5 h) to TH17 (t 5 72 h; Fig. 1c andSupplementary Fig. 1c): early (up to 4 h), intermediate (4–20 h),and late (20–72 h). Each corresponds, respectively, to a differentiationphase5: (1) induction; (2) onset of phenotype and amplification; and(3) stabilization and IL-23 signalling. The early phase is characterized

*These authors contributed equally to this work.

1Broad Institute of MIT and Harvard, 7 Cambridge Center, Cambridge, Massachusetts 02142, USA. 2Center for Neurologic Diseases, Brigham & Women’s Hospital, Harvard Medical School, Boston,Massachusetts 02115, USA. 3Department of Chemistry and Chemical Biology and Department of Physics, Harvard University, Cambridge, Massachusetts 02138, USA. 4Department of Pathology,University of Utah School of Medicine, Salt Lake City, Utah 84132, USA. 5Department of Immunology, St. Jude Children’s Research Hospital, Memphis, Tennessee 38105, USA. 6University of Oxford,Headington Campus, Oxford OX3 7BN, UK. 7Howard Hughes Medical Institute, Department of Biology, Massachusetts Institute of Technology, Cambridge, Massachusetts 02140, USA. {Present address:Translational Health Science & Technology Institute, Faridabad, Haryana 122016, India.

2 5 A P R I L 2 0 1 3 | V O L 4 9 6 | N A T U R E | 4 6 1

Macmillan Publishers Limited. All rights reserved©2013

by transient induction (for example, cluster C5, Fig. 1b) of immuneresponse pathways (for example, IL-6 and TGF-b signalling;Supplementary Table 2). Some early induced genes display sustainedexpression (for example, cluster C10, Fig. 1b); these are enriched fortranscription factors, including the key TH17 factors Stat3, Irf4 andBatf, and the cytokine and cytokine receptors Il21, Lif and Il2ra(Supplementary Table 1). The transition to the intermediate phase(t 5 4 h) is marked by induction of the Rorc gene (encoding the mastertranscription factor ROR-ct; Supplementary Fig. 1d) and another 12transcription factors (cluster C20, Fig. 1b), both known (for example,Ahr) and novel (for example, Trps1) in TH17 differentiation. Duringthe transition to the late phase (t 5 20 h), mRNAs of TH17 signaturecytokines are induced (for example, Il17a, Il9; cluster C19) whereasmRNAs of cytokines that signal other T-cell lineages are repressed(for example, Ifng and Il4). Regulatory cytokines from the IL-10family are also induced (Il10, Il24), possibly as a self-limiting mech-anism related to the emergence of ‘pathogenic’ or ‘non-pathogenic’TH17 cells11. Around 48 h, the cells induce Il23r (data not shown),which has an important role in the late phase (Supplementary Fig. 3and Supplementary Table 1).

Inference of dynamic regulatory interactionsWe proposed the hypothesis that each of the clusters (Fig. 1b andSupplementary Table 2) encompasses genes that share regulatorsactive in the relevant time points. To predict these regulators, weassembled a general network of regulator–target associations frompublished genomics profiles12–19 (Fig. 2a and Methods). We then

connected a regulator to a gene from its set of putative targets onlyif there was also a significant overlap between the regulator’s putativetargets and that gene’s cluster (Methods). Because different regulatorsact at different times, the connection between a regulator and its targetmay be active only within a certain time window. To determine thiswindow, we labelled each edge with a time stamp denoting when boththe target gene is regulated (based on its expression profile) and theregulator node is expressed at sufficient levels (based on its mRNAlevels and inferred protein levels20; Methods). In this way, we derived anetwork ‘snapshot’ for each of the 18 time points (Fig. 2b–d). Overall,10,112 interactions between 71 regulators and 1,283 genes wereinferred in at least one network.

Substantial regulatory re-wiring during differentiationThe active factors and interactions change from one network to thenext. The vast majority of interactions are active only at some timewindows (Fig. 2c), even for regulators (for example, Batf) that par-ticipate in all networks. On the basis of similarity in active interac-tions, we identified three network classes (Fig. 2c) corresponding tothe three differentiation phases (Fig. 2d). We collapsed all networks ineach phase into one model, resulting in three consecutive networkmodels (Fig. 2d, Supplementary Fig. 4 and Supplementary Table 3).Among the regulators, 33 are active in all of the networks (forexample, known master regulators such as Batf, Irf4 and Stat3),whereas 18 are active primarily in one (for example, Stat1 and Irf1in the early network; ROR-ct in the late network). Indeed, whereasRorc mRNA levels are induced at ,4 h, ROR-ct protein levels increase

mRNA

profilingModelling

Perturbation (siRNAs,nanowire delivery)

Validatedmodel

18 time points

(0.5–72 h;

± IL-23: 48–72 h)

3 temporal networks;

71 regulators;

~1,200 target genes

29 transcriptional

regulators;

10 receptors

22 TH17-positive regulators

antagonistically coupled

to 5 TH17-negative regulators

0.5

1 2 4 6 8 10

12

16

20

24

30

48

50

52

42

60

72

0.5

1 2 4 6 8 10

12

16

20

24

30

42

48

50

52

60

72

a

b

C8

C2

C16

C5

C4C10C1C18C3C13C12C14C15

C20C17

C11

C7C6

C19

C9

C5 (n = 112)

F: Immune responseM: Stat1/2, Irf7/9, Il17ra

C15 (n = 119)

F: Immune response

M: Ifng, Ccl3, Tbx21

C20 (n = 31)

F: HaematopoiesisM: Batf, Rorc, Id1/2, Trps1

Rorcinduction

Il17induction

Il23r expression

IL-6 + TGF-β1 TH0

C19 (n = 135)

F: Cytokine signallingM: Il17a, Il10, Il21, Ccr5

Time (0.5–72 h) Time (0.5–72 h)

ClusterTim

e (h)

IL-6 + TGF-β1 (IL-6 + TGF-β1)/(TH0)–3 30

Z score

–2 20

Fold change (log2)

c

0.4 1Correlation (r2)

d

–2 20

Fold change (log2)

Fasl

Xcl1

Cxcl1

0Il6st

Crlf2

Il17ra

Csf2

Nam

pt

Il3

Il15ra

Ccl4

Ccl3

Tnfs

f11

Cm

tm7

Tnfs

f8C

sf1

Eb

i3C

cl1

Ifng

Cd

70

Tnfs

f4Il2

Cxcl1

3T

nfs

f9Ir

ak1b

p1

Ilf3

Clc

f1C

cr6

Are

gC

cr8

Cxcr3

Il2rb

Il16

Cklf

Ccr5

Sp

p1

Il1r1Lif

Lta

Il4ra

Il2ra

Il21

Cxcr4

Inhb

aC

mtm

6N

fil3

Ccr4

Il12rb

2Il24

Il10

Ccl2

0Il1rnIl9

Il17a

0.5

1 2 4 6 8 10

12

16

20

24

30

42

48

50

52

60

72

Tim

e (h)

Tim

e (h)

Figure 1 | Genome-wide temporal expression profiles of TH17differentiation. a, Overview of approach. b, Gene expression profiles duringTH17 differentiation. Shown are the differential expression levels for genes(rows) at 18 time points (columns) in TH17 polarizing conditions (TGF-b1 andIL-6; left panel, Z-normalized per row) or TH17 polarizing conditions relative tocontrol activated TH0 cells (right panel, log2(ratio)). The genes are partitionedinto 20 clusters (C1–C20, colour bars, right). Right: mean expression (y axis)

and standard deviation (error bar) at each time point (x axis) for genes inrepresentative clusters. Cluster size (n), enriched functional annotations (F)and representative genes (M) are denoted. c, Three major transcriptionalphases. Shown is a correlation matrix (red, high; blue, low) between every pairof time points. d, Transcriptional profiles of key cytokines, chemokines andtheir receptors. Time points are the same as in panel b.

RESEARCH ARTICLE

4 6 2 | N A T U R E | V O L 4 9 6 | 2 5 A P R I L 2 0 1 3

Macmillan Publishers Limited. All rights reserved©2013

at approximately 20 h and further rise over time, consistent with ourmodel (Supplementary Fig. 5).

Ranking novel regulators for systematic perturbationIn addition to known TH17 regulators, our network includes dozens ofnovel factors as predicted regulators (Fig. 2d), induced target genes, orboth (Supplementary Fig. 4 and Supplementary Table 3). It also con-tains receptor genes as induced targets, both previously known in TH17cells (for example, Il1r1, Il17ra) and novel (for example, Fas, Itga3).

We ranked candidate regulators for perturbation (Figs 2a and 3a;see Methods), guided by features that reflect a regulatory role (Fig. 3a,‘Network information’) and a role as a target (Fig. 3a, ‘Gene expres-sion information’). We computationally ordered the genes toemphasize certain features (for example, a predicted regulator ofkey TH17 genes) over others (for example, differential expression inour time course data). We used a similar scheme to rank receptor

proteins (Supplementary Table 4 and Methods). Supporting theirquality, our top-ranked factors are enriched (P , 1023) for manuallycurated TH17 regulators (Supplementary Fig. 6), and correlate well(Spearman r . 0.86) with a ranking learned by a supervised method(Methods). We chose 65 genes for perturbation: 52 regulators and 13receptors (Supplementary Table 4). These included most of the top 44regulators and top 9 receptors (excluding a few well known TH17genes and/or those for which knockout data already existed), as wellas additional representative lower ranking factors.

Nanowire-based perturbation of primary T cellsIn unstimulated primary mouse T cells, viral- or transfection-basedsiRNA delivery has been nearly impossible because it either alters differ-entiation or cell viability21,22. We therefore used a new delivery technologybased on silicon nanowires10,23, which we optimized to deliver siRNAeffectively (.95%) into naive T cells without activating them (Fig. 3b, c)23.

b

c

Early Intermediate Late

0.5

1 2 4 6 8 10

12

16

20

24

30

42

48

50

52

60

72

Sta

t1Ir

f1S

tat3

Batf

Irf4

Ro

rcS

marc

a4

Temporal profiles ofindividual interactions

Binding byspecific regulators

0.5 1 2 4 6 8

10

12

16

20

24

30

42

48

50

52

60

72

B C D B C D

A AA

B C D

d

Early Intermediate Late

Late interactions

Sustained interactions

Early interactions

Time (h)

Time course data

Published genomic profiles

Expression clusters

Timing of target gene induction

Presence of regulators

Protein–DNA binding (298 TFs)

DNA binding motifs (823 motifs)

Expression profiles from 159 immune cells

Knockout effects (11 TFs)

NETWORK

MODELIntegrateIntegrate

Ranked list of candidates(TFs and receptors) for

perturbation

Regulator–target

putative associates

Expression changes in key perturbations in TH17 cells

a

Early

transient

interactions

Intermediate

transient

interactions

Sustained

interactions

Late

interactions

0.5 1 2 4 6 8 10

12

16

20

24

30

42

48

50

52

60

72

Early

Inte

rmed

iate

Late

Tbx21Suz12Nr3c1Cdc5lMazIrf3Sp100Cited2Irf7NmiStat1Stat2RelStat4Irf1Irf2Irf9Gata3Zfp281Egr1

Trim24Stat6Satb1CbfbTrp53Hif1aId2Smarca4Irf4BatfSmad2Zfp161Creb1TcfebNcoa3Lrrfip1Stat3Stat5aMycRorcNfil3E2f1Runx2CsdaNfkb1Sap18Zeb2Max

Foxj2Trim28Pou2f2

Sp1RelaRunx1Egr2Ifi35Stat5bE2f4AesLef1Irf8JunEts1Foxo1Fli1Ets2Chd7Phf21aBcl11bEzh1Sp4

0 100

No. of targets (% of max.)

Pert

urb

ed

Kno

wn

No

. o

f ta

rgets

Reg

ula

tor

activity

Time (h)

Figure 2 | A model of the dynamic regulatory network of TH17differentiation. a, Overview of computational analysis. b, Schematic oftemporal network ‘snapshots’. Shown are three consecutive cartoon networks(top and matrix columns), with three possible interactions from regulator (A)to targets (B, C and D), shown as edges (top) and matrix rows (ARB, top row;ARC, middle row; ARD, bottom row). c, Eighteen network ‘snapshots’. Left:each row corresponds to a transcription factor (TF)–target interaction thatoccurs in at least one network; columns correspond to the network at each timepoint. A purple entry an indicates that an interaction is active in that network.The networks are clustered by similarity of active interactions (dendrogram,

top), forming three temporally consecutive clusters (early, intermediate, late;bottom). Right: heat map denoting edges for selected regulators. d, Dynamicregulator activity. Shown is, for each regulator (rows), the number of targetgenes (normalized by its maximum number of targets) in each of the 18networks (columns, left), and in each of the three canonical networks(middle) obtained by collapsing (arrows). Right: regulators chosen forperturbation (pink), known TH17 regulators (grey), and the maximal numberof target genes across the three canonical networks (green, ranging from 0 to250 targets).

ARTICLE RESEARCH

2 5 A P R I L 2 0 1 3 | V O L 4 9 6 | N A T U R E | 4 6 3

Macmillan Publishers Limited. All rights reserved©2013

We attempted to perturb 60 genes with nanowire-mediated siRNAdelivery and achieved efficient knockdown (,60% transcript remain-ing at 48 h post-activation) for 34 genes (Fig. 3c and SupplementaryFig. 7). We obtained knockout mice for seven other genes, two ofwhich (Irf8 and Il17ra) were also in the knockdown set (Sup-plementary Table 4). Altogether, we successfully perturbed 39 ofthe 65 selected genes—29 regulators and 10 receptors—including 21genes not previously associated with TH17 differentiation.

Nanowire-based screen validates 39 network regulatorsWe measured the effects of perturbations at 48 h post-activation onthe expression of 275 signature genes using the Nanostring nCountersystem (Supplementary Tables 5 and 6; Il17ra and Il21r knockoutswere also measured at 60 h). The signature genes were computation-ally chosen to cover as many aspects of the differentiation process aspossible (Methods): they include most differentially expressed cyto-kines, transcription factors, and cell-surface molecules, as well asrepresentatives from each cluster (Fig. 1b) or enriched function (Sup-plementary Table 2), and predicted targets in each network (Sup-plementary Table 3). For validation, we profiled a signature of 86genes using the Fluidigm BioMark system, obtaining highly repro-ducible results (Supplementary Fig. 8).

We scored the statistical significance of a perturbation’s effect on asignature gene by comparing to non-targeting siRNAs and to 18

control genes that were not differentially expressed (SupplementaryInformation and Fig. 4a, all non-grey entries are significant).Supporting the original network model (Fig. 2), there is a significantoverlap between the genes affected by a regulator’s knockdown and itspredicted targets (P # 0.01, permutation test; SupplementaryInformation).

To study the network’s dynamics, we measured the effect of 28 ofthe perturbations at 10 h (shortly after the induction of Rorc;Supplementary Table 5) using the Fluidigm BioMark system. Wefound that 30% of the functional interactions are present with thesame activation/repression logic at both 10 h and 48 h, whereas therest are present only in one time point (Supplementary Fig. 9). This isconsistent with the extent of rewiring in our original model (Fig. 2c).

Two coupled antagonistic circuits in the TH17 networkCharacterizing each regulator by its effect on TH17 signature genes(for example, Il17a, Il17f; Fig. 4b, grey nodes, bottom), we found that,at 48 h, the network is organized into two antagonistic modules: amodule of 22 ‘TH17-positive factors’ (Fig. 4b, blue nodes: 9 novel), theperturbation of which decreased the expression of TH17 signaturegenes (Fig. 4b, grey nodes, bottom), and a module of 5 ‘TH17-negativefactors’ (Fig. 4b, red nodes: 3 novel), the perturbation of which had theopposite effect. Each of the modules is tightly intra-connectedthrough positive, self-reinforcing interactions between its members

b

Candidate regulators (ranked from left to right)

Gene expression information

Network informationa

5 μm

c48-h knockdowns

No

rmaliz

ed

mR

NA

(%

)

50

100

25

75

0

Ccr6

Batf

Etv

6

Eg

r2E

gr1

Arid

5a

Fas

Ets

2

Ccr5

Arn

tl

Rb

pj

Irf4

Po

u2af1

Pro

cr

Min

a

Tg

if1

Id2

Ro

rc

Sm

ox

Trp

s1

Il6st

Trim

24

Pm

l

Irf8

Myd

88

Myc

Sm

arc

a4

Nr3

c1

Ikzf

4

Itg

a3

Gata

3

Kat2

b

Nfk

b1

No

tch1

Tsc22d

3

Sp

4

Zfp

36l1

Zeb

1Z

fp281

Co

ntr

ol

Targets key molecules

Regulator in network

ChIP and knockout

TH17 microarray signature

Differentially expressed

Induced

Expression changes in key

perturbations in TH17 cellsRanking

Low High

Known

Perturbed

Top ranking score

Lowest ranking score

Ro

rcS

atb

1T

rps1

Sm

ox

Ro

raA

rid

5a

Etv

6A

rntl

Ets

1U

be2b

Batf

Sta

t3S

tat1

Sta

t5a

Nr3

c1

Sta

t6T

rim

24

Hif1a

Irf4

Irf8

Ets

2Jun

Runx1

Fli1 Rel

Sp

4E

gr2

Nfk

b1

Zfp

281

Sta

t4R

ela

Tb

x21

Sta

t5b

Irf7

Sta

t2Ir

f3X

bp

1F

oxo

1P

rdm

1A

tf4

Irf1

Gata

3E

gr1

Myc

Cre

b1

Irf9

Irf2

Fo

xj2

Sm

arc

a4

Zfp

161

Trp

53

Su

z12

Po

u2af1

Ceb

pb

Id2

Cre

mM

yst4

Mxi1

Rb

pj

Chd

7C

reb

3l2

Vax2

Klf10

Ski

Elk

3Z

eb

1P

ml

Sert

ad

1N

otc

h1

Lrr

fip

1A

hr

1810007M

14R

IKS

ap

30

Id1

Zfp

238

Vav1

Min

aB

atf

3C

dyl

Ikzf

4N

co

a1

Bcl3

Ju

nb

Ss18

Phf1

3M

ta3

Asxl1

Lass4

Skil

Dd

it3

Fo

sl2

Runx2

Tle

1A

tf3

Ell2

Aes

Bcl1

1b

Jarid

2K

lf9

Kat2

bK

lf6

E2f8

Bcl6

Znrf

2T

sc22d

3K

lf7

Hm

gb

2F

us

Sirt2

Maff

Chm

p1b

Gata

d2b

Sm

ad

7Z

fp703

Znrf

1Jm

jd1c

Zfp

36l2

Tsc22d

4N

fe2l2

Rnf1

1A

rid

3a

Men1

Rara

Cb

x4

Zfp

62

Cic

Hcls

1Z

fp36l1

Tg

if1

Figure 3 | Knockdown screen in TH17 differentiation using siliconnanowires. a, Unbiased ranking of perturbation candidates. Shown are thegenes ordered from left to right based on their ranking for perturbation(columns, top ranking is left-most). Two top matrices: criteria for ranking by‘Network information’ (topmost) and ‘Gene expression information’. Purpleentry: gene has the feature (intensity proportional to feature strength; top fivefeatures are binary). Bar chart indicates ranking score. ‘Perturbed’ row: darkgrey, genes successfully perturbed by knockdown followed by high-qualitymRNA quantification; light grey, genes that we attempted to knockdown butcould not achieve or maintain sufficient knockdown or did not obtain enough

replicates; Black, genes that we perturbed by knockout or for which knockoutdata were already available. Known row: orange entry: a gene was previouslyassociated with TH17 function (this information was not used to rank the genes;Supplementary Fig. 6). b, Scanning electron micrograph of primary T cells(false-coloured purple) cultured on vertical silicon nanowires. c, Effectiveknockdown by siRNA delivered on nanowires. Shown is the percentage ofmRNA remaining after knockdown (by qPCR, y axis: mean 6 standard errorrelative to non-targeting siRNA control, n 5 12, black bar on left) at 48 h afteractivation.

RESEARCH ARTICLE

4 6 4 | N A T U R E | V O L 4 9 6 | 2 5 A P R I L 2 0 1 3

Macmillan Publishers Limited. All rights reserved©2013

(70% of the intra-module edges), whereas most (88%) inter-moduleinteractions are negative. This organization, which is statistically sig-nificant (empirical P value ,1023; Methods, Supplementary Fig. 10),is reminiscent of that observed previously in genetic circuits inyeast24,25. At 10 h, the same regulators do not yield this clear pattern(P . 0.5), suggesting that, at that point, the network is still malleable.

The two antagonistic modules may have a key role in maintaining thebalance between TH17 and other T-cell subsets and in self-limiting thepro-inflammatory status of TH17 cells. Indeed, perturbing TH17-positive

factors also induces signature genes of other T-cell subsets, whereasperturbing TH17-negative factors suppresses them (for example, Foxp3,Gata3 and Stat4; Fig. 4b, grey nodes, top).

Validation and characterization of novel factorsNext, we focused on the role of 12 of the positive or negative factors(including 11 of the 12 novel factors that have not been associatedwith TH17 cells; Fig. 4b, light-grey halos). After knockdown of eachfactor, we used RNA-seq analysis to test whether its predicted targets

a bIt

ga3

Pert

urb

ed

TH17 (negative)TH17 (positive)N

r3c1

Etv

6F

as

Myc

Batf

Eg

r2Il17ra

Irf4

(K

O)

Ccr6

Pro

cr

Irf8

No

tch1

Min

aE

gr1

Zeb

1N

fkb

1P

ml

Sm

arc

a4

Po

u2af1

Rb

pj

Il17ra

(K

O, 60 H

)Il21r

(KO

, 60 H

)Il21r

(KO

)Ir

f8 (K

O)

Tg

if1

Trp

s1

Id2

Sm

ox

Trim

24

Il6st

Zfp

36l1

Il17ra

(K

O)

Ccr5

Kat2

bM

yd

88

Il12rb

1 (K

O)

Sta

t1 (K

O)

Ets

2Ik

zf4

Sp

4T

sc22d

3Ir

f1 (K

O)

Il17, Il21, Il10, Il24, Il9, Lif, Il22, Rorc, Batf, Irf4, Mina, Ahr, Rbpj, Procr

Foxp3, Tsc22d3, Ikzf4, Il7r, Serpinb1a, Il10ra

Maff, Egr2, Ccl20

Gata3, Stat4, Irf8

‘Positive’ module

–2 20

Fold change (log2)

c d

Fold enrichment of upregulation (log2)

–0.50 0.50

14 15 9 20 19 12 7 13 11 3 10 4 1 17 6 18 5 2 16 8

Pou2af1Etv6Mina

Irf8Smarca4

Egr2Procr

FasZeb1

Tsc22d3Sp4

Ikzf4

Cluster

0

‘Positive’module

‘Negative’module

Signature of other CD4+ T cells

RorcIl21Il17aIl17f Stat3

Signature of TH17

‘Negative’module

Downregulation

Upregulation

IL-17 downregulated in KD

IL-17 upregulated in KD

Novel factors

Myc

Irf8

Il21r

Mina

Il17ra

Irf4

Irf1

Sp4

Tbx21Ifng Stat4Gata3

Ikzf4

Ets2

Foxp3

Tsc22d3

Notch1Pml

Rbpj

Zeb1

Batf

Pou2af1

Procr

Smarca4

Nfkb1

Ccr5

Ccr6

Egr1

Egr2

Fas Etv6

Il12rb1

P v

alu

e (–lo

g10)

0

5

10

15

20

All

Early

Inte

rmed

iate

Late

Sp4

All

Early

Inte

rmed

iate

Late

Smarca4

All

Early

Inte

rmed

iate

Late

Egr2

All

Inte

rmed

iate

Irf8

Late

Predicted targets(network)

Affected genes(knockdown)

P value

Genes in temporal cluster

Affected genes(knockdown)

P value,up- or downregulation

Figure 4 | Coupled and mutually antagonistic modules in the TH17network. a, Impact of perturbed genes on a 275-gene signature. Shown arechanges in the expression of 275 signature genes (rows) following knockdown orknockout (KO) of 39 factors (columns) at 48 h (as well as Il21r and Il17raknockout at 60 h). Blue, decreased expression of target following perturbation of aregulator (compared to a non-targeting control); red, increased expression; grey,not significant; all non-grey entries are significant (Supplementary Information).Perturbed (left): signature genes that are also perturbed as regulators (blackentries). Key signature genes are denoted on right. b, Two coupled and opposingmodules. Shown is the perturbation network associating the ‘positive regulators’(blue nodes) of TH17 signature genes, the ‘negative regulators’ (red nodes), TH17signature genes (grey nodes, bottom) and signature genes of other CD41 T cells(grey nodes, top). A blue edge from node A to B indicates that knockdown of Adownregulates B; a red edge indicates that knockdown of A upregulates B. Light-grey halos: regulators not previously associated with TH17 differentiation.c, Knockdown effects validate edges in network model. Venn diagram: we

compare the set of targets for a factor in the original model of Fig. 2a (pink circle)to the set of genes that respond to that factor’s knockdown in an RNA-seqexperiment (yellow circle). Bar chart (bottom): shown is the 2log10(P value) (yaxis, hypergeometric test) for the significance of this overlap for four factors (xaxis). Similar results were obtained with a non-parametric rank-sum test (Mann–Whitney U-test, Supplementary Information). Red dashed line: P 5 0.01.d, Global knockdown effects are consistent across clusters. Venn diagram: wecompare the set of genes that respond to a factor’s knockdown in an RNA-seqexperiment (yellow circle) to each of the 20 clusters of Fig. 1b (purple circle). Weexpect the knockdown of a ‘TH17 positive’ regulator to repress genes in inducedclusters, and induce genes in repressed clusters (and vice versa for ‘TH17 negative’regulators). Heat map: for each regulator knockdown (rows) and each cluster(columns) shown are the significant overlaps (non-grey entries) by the test above.Red, fold enrichment for upregulation upon knockdown; blue, fold enrichmentfor downregulation upon knockdown. Orange entries in the top row indicateinduced clusters.

ARTICLE RESEARCH

2 5 A P R I L 2 0 1 3 | V O L 4 9 6 | N A T U R E | 4 6 5

Macmillan Publishers Limited. All rights reserved©2013

(Fig. 2) were affected (Fig. 4c, Venn diagram, top). We found highlysignificant overlaps (P , 1025) for three of the factors (Egr2, Irf8 andSp4) that exist in both data sets, and a borderline significant overlapfor the fourth (Smarca4), validating the quality of the edges in ournetwork.

Next, we assessed the designation of each of the 12 factors as ‘TH17positive’ or ‘TH17 negative’ by comparing the set of genes that respondto that factor’s knockdown (in RNA-seq) to each of the 20 clusters(Fig. 1b). Consistent with the original definitions, knockdown of aTH17-positive regulator downregulated genes in otherwise inducedclusters and upregulated genes in otherwise repressed or uninducedclusters (and vice versa for TH17-negative regulators; Fig. 4d andSupplementary Fig. 11a, b). The genes affected by either positive ornegative regulators also significantly overlap with those bound by keyCD41 transcription factors (for example, Foxp3 (refs 26, 27), Batf,Irf4 and ROR-ct (refs 28, 29), S. Xiao et al., unpublished data).

Mina promotes the TH17 and inhibits the Foxp3 programKnockdown of Mina, a chromatin regulator from the Jumonji C(JmjC) family, represses the expression of signature TH17 cytokinesand transcription factors (for example, Rorc, Batf, Irf4) and of late-induced genes (clusters C9 and C19; P , 1025; Supplementary Tables5 and 7), while increasing the expression of Foxp3, the master trans-cription factor of Treg cells. Mina is strongly induced during TH17differentiation (cluster C7), is downregulated in Il23r2/2 TH17 cells,and is a predicted target of Batf30, ROR-ct30 and Myc in our model(Fig. 5a). Mina was shown to suppress TH2 bias by interacting with thetranscription factor NFAT and repressing the Il4 promoter31. How-ever, in our cells, Mina knockdown did not induce TH2 genes, indi-cating an alternative mode of action via positive feedback loopsbetween Mina, Batf and ROR-ct (Fig. 5a, left). Consistent with thismodel, Mina expression is reduced in TH17 cells from Rorc knockoutmice, and the Mina promoter was found to be bound by ROR-ctby ChIP-seq (data not shown). Finally, the genes induced by Minaknockdown significantly overlap with those bound by Foxp3 in Treg

cells26,27 (P , 10225; Supplementary Table 7) and with a cluster previ-ously linked to Foxp3 activity in Treg cells32 (Supplementary Fig. 11cand Supplementary Table 7).

To analyse the role of Mina further, we measured IL-17a and Foxp3expression after differentiation of naive T cells from Mina2/2 mice.Mina2/2 cells had decreased IL-17a and increased Foxp31 T cellscompared to wild-type cells, as detected by intracellular staining(Fig. 5a). Cytokine analysis of the corresponding supernatants con-firmed a decrease in IL-17a production and an increase in IFN-c(Fig. 5a) and TNF-a (Supplementary Fig. 12a). This is consistent witha model where Mina, induced by ROR-ct and Batf, promotes transcrip-tion of Rorc, while suppressing induction of Foxp3, thus affecting thereciprocal Treg/TH17 balance33 by favouring rapid TH17 differentiation.

Fas promotes the TH17 program and suppresses IFN-cFas, the TNF receptor superfamily member 6, is another TH17-positiveregulator (Fig. 5b). Fas is induced early and is a target of Stat3 and Batfin our model. Fas knockdown represses the expression of key TH17genes (for example, Il17a, Il17f, Hif1a, Irf4 and Rbpj) and of theinduced cluster C14, and promotes the expression of TH1-relatedgenes, including Ifngr1 and Klrd1 (CD94; by RNA-seq, Figs 4, 5b,Supplementary Table 7 and Supplementary Fig. 11). Fas- and Fas-ligand-deficient mice are resistant to the induction of autoimmuneencephalomyelitis (EAE)34, but have no defect in IFN-c or TH1 res-ponses. The mechanism underlying this phenomenon has not beenidentified.

To explore this, we differentiated T cells from Fas2/2 mice (Fig. 5band Supplementary Fig. 12c). Consistent with our knockdown ana-lysis, expression of IL-17a was strongly repressed and IFN-c produc-tion was strongly increased under both TH17 and TH0 polarizingconditions (Fig. 5b). These results suggest that besides being a deathreceptor, Fas may have an important role in controlling the TH1/TH17balance, and Fas2/2 mice may be resistant to EAE due to lack of TH17cells.

Pou2af1 promotes the TH17 program and suppresses IL-2expressionKnockdown of Pou2af1 (also called OBF1) strongly decreases theexpression of TH17 signature genes (Fig. 5c) and of intermediate-and late-induced genes (clusters C19 and C20, P , 1027; Supplemen-tary Tables 5 and 7), while increasing the expression of regulators ofother CD41 subsets (for example, Foxp3, Stat4, Gata3) and of genes in

a

Rorc

Foxp3

IL-23r

IL-17a

Myc

Batf

Mina

b

IL-17a

Fas

IL-23rBatfStat3

Rbpj

No

cyto

kin

es

TG

F-β

1+

IL-6

WT Fas–/–

3.04 0.15

0.3796.4

0.27 1.26

39.459

16.2 0.32

83.2 0.3

1.12 1.85

24.472.6

IFN

-γ

IL-17a

0

1,000

2,000

3,000 *** ***

0

100

200

300

400

***

d

IL-21Foxp3 Mt1AhrCebpb

Batf IL-17a Irf4

Tsc22d3

Rorc

IL-17 module

IL-10 module

Downregulation

Upregulation

Only binding

Only perturbation effect

Pivot node

Downregulated in KD of pivot nodes

Upregulated in KD of pivot nodes

0

10

20

30

40

50

60

70

Overlap

(%

)

Sta

t3

Sta

t5

Fo

xp

3 (T

reg)

Ro

rc

Irf4

Batf

Genomic regions Bound genes

Batf

Pou2af1

RorcStat4Gata3

Stat3

Irf4IL-17aFoxp3

c

TG

F-β

1+

IL-6

No

cyto

kin

es

IL-2

WT Pou2af1–/–

IL-17a

13.2 0.083

0.4586.2

0.1 0.058

10.189.7

0.81 4.35 × 10–3

0.2299

0.55 0.31

1881.2

0

50

100

150 ***IL

-17 (ng

ml–

1)

Mina–/–

0.045 2.99 × 10–3

8.3891.6

0.021 6.02 × 10–3

19.880.1

WT

CD

4T

GF

-β1+

IL-6

TG

F-β

1+

IL-6

Foxp3

IFN

-γ

IL-17a

0

25

50

75

100

125

150

175***

0204060

500

1,000

1,500*

2.39 10

IFN

-γ (n

g m

l–1)

IL-1

7 (ng

ml–

1)

*****

* P < 0.05

P < 0.01

P < 0.001

Mina

KO T H

17

WT

T H17

Mina

KO T H

0

WT

T H0

Mina

KO T H

17

WT

T H17

Mina

KO T H

0

WT

T H0

Fas KO

T H17

WT

T H17

Fas KO

T H0

WT

T H0

Fas KO

T H17

WT

T H17

Fas KO

T H0

WT

T H0

Pou2a

f1 K

O T H

17

WT

T H17

Pou2a

f1 K

O T H

0

WT

T H0

IFN

-γ (n

g m

l–1)

IL-1

7 (ng

ml–

1)

Figure 5 | Mina, Fas, Pou2af1 and Tsc22d3 are key novel regulatorsaffecting the TH17 differentiation programs. a–d, Left: shown are regulatorynetwork models centred on different pivotal regulators (square nodes; a, Mina;b, Fas; c, Pou2af1; d, Tsc22d3). In each network, shown are the targets andregulators (round nodes) connected to the pivotal nodes based on perturbation(red and blue dashed edges), transcription factor binding (black solid edges), orboth (red and blue solid edges). Genes affected by perturbing the pivotal nodesare coloured (blue, target is downregulated by knockdown of pivotal node; red,target is upregulated). Middle and right panels of a–c: intracellular staining and

cytokine assays by ELISA or cytometric bead assays (CBA) on culturesupernatants at 72 h of in vitro differentiated cells from respective knockoutmice activated in vitro with anti-CD3 plus anti-CD28 with or without TH17polarizing cytokines (TGF-b1 plus IL-6). d, Middle: ChIP-seq of Tsc22d3.Shown is the proportion of overlap in bound genes (dark grey) or boundregions (light grey) between Tsc22d3 and a host of TH17 canonical factors (xaxis). All results are statistically significant (P , 1026; Hypergeometric score(gene overlap) and Binomial score (region overlap); SupplementaryInformation).

RESEARCH ARTICLE

4 6 6 | N A T U R E | V O L 4 9 6 | 2 5 A P R I L 2 0 1 3

Macmillan Publishers Limited. All rights reserved©2013

non-induced clusters (clusters C2 and C16, P , 1029; SupplementaryTable 5 and 7). The role of Pou2af1 in T-cell differentiation has notbeen explored35.

To investigate its effects, we differentiated T cells from Pou2af12/2

mice (Fig. 5c and Supplementary Fig. 12b). Compared to wild-typecells, IL-17a production was strongly repressed. Interestingly, IL-2production was strongly increased in Pou2af12/2 T cells undernon-polarizing (TH0) conditions. Thus, Pou2af1 may promoteTH17 differentiation by blocking production of IL-2, a known endo-genous repressor of TH17 cells36. Pou2af1 acts as a transcriptional co-activator of the transcription factors Oct1 or Oct2 (ref. 35). IL-17aproduction was also strongly repressed in Oct1-deficient cells(Supplementary Fig. 12d), suggesting that Pou2af1 may exert someof its effects through this co-factor.

Tsc22d3 may limit TH17 generation and inflammationKnockdown of the TSC22 domain family protein 3 (Tsc22d3)increases the expression of TH17 cytokines (Il17a, Il21) and transcrip-tion factors (Rorc, Rbpj, Batf), and reduces Foxp3 expression. Previousstudies in macrophages have shown that Tsc22d3 expression is stimu-lated by glucocorticoids and IL-10, and it has a key role in their anti-inflammatory and immunosuppressive effects37. Tsc22d3 knockdownin TH17 cells increased the expression of Il10 and other key genes thatenhance its production (Fig. 5d). Although IL-10 production has beenshown33,38,39 to render TH17 cells less pathogenic in autoimmunity,co-production of IL-10 and IL-17a may be the indicated response forclearing certain infections such as Staphylococcus aureus at mucosalsites40. This suggests a model where Tsc22d3 is part of a negativefeedback loop for the induction of a TH17 cell subtype that co-produces IL-17 and IL-10 and limits their pro-inflammatory capa-city. Tsc22d3 is induced in other cells in response to the steroiddexamethasone41, which represses TH17 differentiation and Rorcexpression42. Thus, Tsc22d3 may mediate this effect of steroids.

To characterize the role of Tsc22d3 further, we used ChIP-seq tomeasure its DNA-binding profile in TH17 cells and RNA-seq follow-ing its knockdown to measure its functional effects. There is a signifi-cant overlap between Tsc22d3’s functional and physical targets(P , 0.01, for example, Il21, Irf4; Supplementary Information andSupplementary Table 8). For example, Tsc22d3 binds to Il21 andIrf4, which also become upregulated in the Tsc22d3 knockdown.Furthermore, the Tsc22d3 binding sites significantly overlap thoseof major TH17 factors, including Batf, Stat3, Irf4 and ROR-ct (.5-foldenrichment; Fig. 5d, Supplementary Table 8 and SupplementaryMethods). This suggests a model where Tsc22d3 exerts its TH17-nega-tive function as a transcriptional repressor that competes with TH17-positive regulators over binding sites, analogous to previous findings inCD41 regulation29,43.

DiscussionWe combined a high-resolution transcriptional time course, novelmethods to reconstruct regulatory networks, and innovative nano-technology to perturb T cells, to construct and validate a networkmodel for TH17 differentiation. The model consists of three consec-utive, densely intra-connected networks, implicates 71 regulators (46novel), and suggests substantial rewiring in 3 phases. The 71 regula-tors significantly overlap with genes genetically associated withinflammatory bowel disease44 (11 of 71, P , 1029). Building on thismodel, we systematically ranked 127 putative regulators (82 novel;Supplementary Table 4) and tested top ranking ones experimentally.

We found that the TH17 regulators are organized into two tightlycoupled, self-reinforcing but mutually antagonistic modules, thecoordinated action of which may explain how the balance betweenTH17, Treg and other effector T-cell subsets is maintained, and howprogressive directional differentiation of TH17 cells is achieved.Within the two modules are 12 novel factors (Figs 4 and 5), which

we further characterized, highlighting four of the factors (others are inSupplementary Note and Supplementary Fig. 13).

A recent study29 systematically ranked TH17 regulators based onChIP-seq data for known key factors and transcriptional profiles inwild-type and knockout cells. Whereas their network centred onknown core TH17 transcription factors, our complementary approachperturbed many genes in a physiologically meaningful setting.Reassuringly, their core TH17 network significantly overlaps withour computationally inferred model (Supplementary Fig. 14).

The wiring of the positive and negative modules (Figs 4 and 5)uncovers some of the functional logic of the TH17 program, but pro-bably involves both direct and indirect interactions. Our functionalmodel provides an excellent starting point for deciphering the under-lying physical interactions with DNA binding profiles30 or protein–protein interactions (accompanying paper45). The regulators that weidentified are compelling new targets for regulating the TH17/Treg

balance and for switching pathogenic TH17 into non-pathogenic ones.

METHODS SUMMARYWe measured gene expression profiles at 18 time points (0.5 to 72 h) under TH17conditions (IL-6, TGF-b1) or control (TH0) using Affymetrix microarraysHT_MG-430A. We detected differentially expressed genes using a consensusover four inference methods, and clustered the genes using k-means clustering,with an automatically derived k. Temporal regulatory interactions were inferredby looking for significant (P , 5 3 1025 and fold enrichment .1.5) overlapsbetween the regulator’s putative targets (for example, based on ChIP-seq) andthe target gene’s cluster (using four clustering schemes). Candidates for perturba-tion were ordered lexicographically using network-based and expression-basedfeatures. Perturbations were done using SiNW for siRNA delivery.

Full Methods and any associated references are available in the online version ofthe paper.

Received 28 September 2012; accepted 5 February 2013.

Published online 6 March 2013.

1. Bettelli, E., Oukka, M. & Kuchroo, V. K. TH-17 cells in the circle of immunity andautoimmunity. Nature Immunol. 8, 345–350 (2007).

2. Zhou, L. et al. TGF-b-induced Foxp3 inhibits TH17 cell differentiation byantagonizing RORct function. Nature 453, 236–240 (2008).

3. O’Shea, J. et al. Signal transduction and Th17 cell differentiation. Microbes Infect.11, 599–611 (2009).

4. Zhou, L. & Littman, D. Transcriptional regulatory networks in Th17 celldifferentiation. Curr. Opin. Immunol. 21, 146–152 (2009).

5. Korn, T., Bettelli, E., Oukka, M. & Kuchroo, V. K. IL-17 and Th17 cells. Annu. Rev.Immunol. 27, 485–517 (2009).

6. Amit, I. et al. Unbiased reconstruction of a mammalian transcriptional networkmediating pathogen responses. Science 326, 257–263 (2009).

7. Novershtern, N. et al. Densely interconnected transcriptional circuits control cellstates in human hematopoiesis. Cell 144, 296–309 (2011).

8. Litvak, V. et al. Function of C/EBPd in a regulatory circuit that discriminatesbetween transient and persistent TLR4-induced signals. Nature Immunol. 10,437–443 (2009).

9. Suzuki, H. et al. The transcriptional network that controls growth arrest anddifferentiation in a human myeloid leukemia cell line. Nature Genet. 41, 553–562(2009).

10. Shalek, A. K. et al. Vertical silicon nanowires as a universal platform for deliveringbiomolecules into living cells. Proc. Natl Acad. Sci. USA 107, 1870–1875 (2010).

11. Lee, Y. et al. Induction and molecular signature of pathogenic TH17 cells. NatureImmunol. 13, 991–999 (2012).

12. Linhart, C., Halperin, Y. & Shamir, R. Transcription factor and microRNA motifdiscovery: the Amadeus platform and a compendium of metazoan target sets.Genome Res. 18, 1180–1189 (2008).

13. Zheng, G. et al. ITFP: an integrated platform of mammalian transcription factors.Bioinformatics 24, 2416–2417 (2008).

14. Wilson, N. K. et al. Combinatorial transcriptional control in blood stem/progenitorcells: genome-wide analysis of ten major transcriptional regulators. Cell Stem Cell7, 532–544 (2010).

15. Lachmann, A. et al. ChEA: transcription factor regulation inferred from integratinggenome-wide ChIP-X experiments. Bioinformatics 26, 2438–2444 (2010).

16. Liberzon, A. et al. Molecular signatures database (MSigDB) 3.0. Bioinformatics 27,1739–1740 (2011).

17. Jiang, C., Xuan, Z., Zhao, F. & Zhang, M. TRED: a transcriptional regulatory elementdatabase, new entries and other development. Nucleic Acids Res. 35, D137–D140(2007).

18. Elkon, R., Linhart, C., Sharan, R., Shamir, R. & Shiloh, Y. Genome-wide in silicoidentification of transcriptional regulators controlling the cell cycle in human cells.Genome Res. 13, 773–780 (2003).

ARTICLE RESEARCH

2 5 A P R I L 2 0 1 3 | V O L 4 9 6 | N A T U R E | 4 6 7

Macmillan Publishers Limited. All rights reserved©2013

19. Heng, T. S. & Painter, M. W. The Immunological Genome Project: networks of geneexpression in immune cells. Nature Immunol. 9, 1091–1094 (2008).

20. Schwanhausser, B. et al. Global quantification of mammalian gene expressioncontrol. Nature 473, 337–342 (2011).

21. Dardalhon, V. et al. Lentivirus-mediated gene transfer in primary T cells isenhanced by a central DNA flap. Gene Ther. 8, 190–198 (2001.

22. McManus, M. et al. Small interfering RNA-mediated gene silencing in Tlymphocytes. J. Immunol. 169, 5754 (2002).

23. Shalek, A. K. et al. Nanowire-mediated delivery enables functional interrogation ofprimary immune cells: application to the analysis of chronic lymphocyticleukemia. Nano Lett. 12, 6498–6504 (2012).

24. Segre, D., Deluna, A., Church, G. M. & Kishony, R. Modular epistasis in yeastmetabolism. Nature Genet. 37, 77–83 (2005).

25. Peleg, T., Yosef, N., Ruppin, E. & Sharan, R. Network-free inference of knockouteffects in yeast. PLOS Comput. Biol. 6, e1000635 (2010).

26. Marson, A. et al. Foxp3 occupancy and regulation of key target genes during T-cellstimulation. Nature 445, 931–935 (2007).

27. Zheng, Y. et al. Genome-wide analysis of Foxp3 target genes in developing andmature regulatory T cells. Nature 445, 936–940 (2007).

28. Glasmacher, E. et al. A genomic regulatory element that directs assembly andfunction of immune-specific AP-1-IRF complexes. Science 338, 975–980 (2012).

29. Ciofani,M.et al. A validated regulatory network for Th17cell specification.Cell 151,289–303 (2012).

30. Glasmacher, E. et al. A genomic regulatory element that directs assembly andfunction of immune-specific AP-1-IRF complexes. Science 338, 975–980 (2012).

31. Okamoto, M. et al. Mina, an Il4 repressor, controls T helper type 2 bias. NatureImmunol. 10, 872–879 (2009).

32. Hill, J. A. et al. Foxp3 transcription-factor-dependent and -independent regulationof the regulatory T cell transcriptional signature. Immunity 27, 786–800 (2007).

33. Korn, T. et al. IL-21 initiates an alternative pathway to induce proinflammatoryTH17 cells. Nature 448, 484–487 (2007).

34. Waldner, H., Sobel, R. A., Howard, E. & Kuchroo, V. K. Fas- and FasL-deficient miceare resistant to induction of autoimmune encephalomyelitis. J. Immunol. 159,3100–3103 (1997).

35. Teitell, M. A. OCA-B regulation of B-cell development and function. TrendsImmunol. 24, 546–553 (2003).

36. Laurence, A. et al. Interleukin-2 signaling via STAT5 constrains T helper 17 cellgeneration. Immunity 26, 371–381 (2007).

37. Choi, S.-J. et al. Tsc-22 enhances TGF-b signaling by associating with Smad4 andinduces erythroid cell differentiation. Mol. Cell. Biochem. 271, 23–28 (2005).

38. Peters, A., Lee, Y.&Kuchroo,V.K.Themany facesofTh17cells.Curr. Opin. Immunol.23, 702–706 (2011).

39. Chaudhry, A. et al. Interleukin-10 signaling in regulatory T cells is required forsuppression of Th17 cell-mediated inflammation. Immunity 34, 566–578 (2011).

40. Zielinski, C. E. et al. Pathogen-induced human TH17 cells produce IFN-c or IL-10and are regulated by IL-1b. Nature 484, 514–518 (2012).

41. Jing, Y. et al. A mechanistic study on the effect of dexamethasone in moderatingcell death in Chinese Hamster Ovary cell cultures. Biotechnol. Prog. 28, 490–496(2012).

42. Hu, S. M., Luo, Y. L., Lai, W. Y. & Chen,P. F. Effectsofdexamethasone on intracellularexpression of Th17 cytokine interleukin 17 in asthmatic mice [in Chinese]. NanFang Yi Ke Da Xue Xue Bao 29, 1185–1188 (2009).

43. Yang, X. P. et al. Opposing regulation of the locus encoding IL-17 throughdirect, reciprocal actions of STAT3 and STAT5. Nature Immunol. 12, 247–254(2011).

44. Jostins, L. et al. Host–microbe interactions have shaped the genetic architecture ofinflammatory bowel disease. Nature 491, 119–124 (2012).

45. Wu, C. et al. Induction of pathogenic TH17 cells by inducible salt-sensing kinaseSGK1. Nature http://dx.doi.org/10.1038/nature11984 (2013).

Supplementary Information is available in the online version of the paper.

Acknowledgements We thank L. Gaffney and L. Solomon for artwork, the Broad’sGenomics Platform for sequencing, and D. Kozoriz for cell sorting. Work was supportedby NHGRI (1P50HG006193-01 to H.P. and A.R.), NIH Pioneer Awards(5DP1OD003893-03 to H.P., DP1OD003958-01 to A.R.), NIH (NS 30843, NS045937,AI073748 and AI45757 to V.K.K.), National MS Society (RG2571 to V.K.K.), HHMI (A.R.),and the Klarman Cell Observatory (A.R.).

Author Contributions N.Y., A.K.S., J.T.G., H.P., V.K.K. and A.R. conceived the study anddesigned experiments. N.Y. developed computational methods. N.Y., A.K.S. and J.T.G.analysed the data. A.K.S., J.T.G., H.J., Y.L., A.A., C.W., K.K., S.X., M.J., D.G., R.S., D.Y.L. andJ.J.T. conducted the experiments. A.S., M.R.P., P.J.R., M.L.C., M.B. and D.T. providedknockoutmice.N.Y., A.K.S., J.T.G., V.K.K.,H.P. andA.R.wrote thepaperwith input fromallthe authors.

Author Information The microarray, RNA-seq and ChIP-seq data sets have beendeposited in the Gene Expression Omnibus database under accession numbersGSE43955, GSE43969, GSE43948 and GSE43949. Reprints and permissionsinformation is available at www.nature.com/reprints. The authors declare nocompeting financial interests. Readers are welcome to comment on the online versionof the paper. Correspondence and requests for materials should be addressed to H.P.(Hongkun_Park@harvard.edu), V.K.K. (vkuchroo@rics.bwh.harvard.edu) or A.R.(aregev@broad.mit.edu).

RESEARCH ARTICLE

4 6 8 | N A T U R E | V O L 4 9 6 | 2 5 A P R I L 2 0 1 3

Macmillan Publishers Limited. All rights reserved©2013

METHODSMice. C57BL/6 wild-type, Irf12/2, Fas2/2, Irf4fl/fl and Cd4Cre mice were obtainedfrom Jackson Laboratory. Stat12/2 and 129/Sv control mice were purchased fromTaconic. Il12rb12/2 mice were provided by P. Kalipada. Il17ra2/2 mice wereprovided by J. Kolls. Irf8fl/fl mice were provided by K. Ozato. Both Irf4fl/fl andIrf8fl/fl mice were crossed to Cd4Cre mice to generate Cd4Cre3Irf4fl/fl andCd4Cre3Irf8fl/fl mice. All animals were housed and maintained in a conventionalpathogen-free facility at the Harvard Institute of Medicine in Boston (IUCACprotocols: 0311-031-14 (V.K.K.) and 0609-058015 (A.R.)). All experiments wereperformed in accordance to the guidelines outlined by the Harvard Medical AreaStanding Committee on Animals at the Harvard Medical School. In addition,spleens from Mina2/2 mice were provided by M. Bix (IACUC protocol: 453).Pou2af12/2 mice were obtained from the laboratory of R. Roeder46. Wild-typeand Oct12/2 fetal livers were obtained at day E12.5 and transplanted into sub-lethally irradiated Rag12/2 mice as previously described47 (IACUC protocol: 11-09003).Cell sorting and in vitro T-cell differentiation. CD41 T cells were purified fromspleen and lymph nodes using anti-CD4 microbeads (Miltenyi Biotech) thenstained in PBS with 1% FCS for 20 min at room temperature with anti-CD4-PerCP, anti-CD62l-APC and anti-CD44-PE antibodies (all Biolegend). NaiveCD41CD62lhighCD44low T cells were sorted using the BD FACSAria cell sorter.Sorted cells were activated with plate-bound anti-CD3 (2mg ml21) and anti-CD28 (2 mg ml21) in the presence of cytokines. For TH17 differentiation: 2 ngml21 rhTGF-b1 (Miltenyi Biotec), 25 ng ml21 rmIl-6 (Miltenyi Biotec), 20 ngml21 rmIl-23 (Miltenyi Biotec), and 20 ng ml21 rmIL-b1 (Miltenyi Biotec).Cells were cultured for 0.5–72 h and collected for RNA, intracellular cytokinestaining, and flow cytometry.Flow cytometry and intracellular cytokine staining. Sorted naive T cells werestimulated with phorbol 12-myristate 13-aceate (PMA) (50 ng ml21, Sigma-aldrich), ionomycin (1mg ml21, Sigma-aldrich) and a protein transport inhibitorcontaining monensin (Golgistop) (BD Biosciences) for 4 h before detection bystaining with antibodies. Surface markers were stained in PBS with 1% FCS for20 min at room temperature, then subsequently the cells were fixed in Cytoperm/Cytofix (BD Biosciences), permeabilized with Perm/Wash Buffer (BD Biosciences)and stained with Biolegend conjugated antibodies, that is, Brilliant violet 650 anti-mouse IFN-c (XMG1.2) and allophycocyanin-anti-IL-17A (TC11-18H10.1),diluted in Perm/Wash buffer as described48 (Fig. 5 and Supplementary Fig. 11).To measure the time course of ROR-ct protein expression, a phycoerythrin-conjugated anti- retinoid-related orphan receptor-c was used (B2D), also fromeBioscience (Supplementary Fig. 4). Foxp3 staining for cells from knockout micewas performed with the Foxp3 staining kit by eBioscience (00-5523-00) in accord-ance with their ‘One-step protocol for intracellular (nuclear) proteins’. Data werecollected using either a FACS Calibur or LSR II (Both BD Biosciences), thenanalysed using Flow Jo software (Treestar)49,50.Quantification of cytokine secretion using ELISA. Naive T cells from knockoutmice and their wild-type controls were cultured as described above, their super-natants were collected after 72 h, and cytokine concentrations were determinedby ELISA (antibodies for IL-17 and IL-10 from BD Bioscience) or by cytometricbead array for the indicated cytokines (BD Bioscience), according to the manu-facturers’ instructions (Fig. 5 and Supplementary Fig. 11).Microarray data. Naive T cells were isolated from wild-type mice, and treatedwith IL-6 and TGF-b1. Affymetrix microarrays HT_MG-430A were used tomeasure the resulting mRNA levels at 18 different time points (0.5–72 h;Fig. 1b). Cells treated initially with IL-6, TGF-b1 and with addition of IL-23 after48 h were profiled at four time points (50–72 h). As control, we used time- andculture-matched wild-type naive T cells stimulated under TH0 conditions.Biological replicates were measured in 8 of the 18 time points (1 h, 4 h, 10 h,20 h, 30 h, 42 h, 52 h, 60 h) with high reproducibility (r2 . 0.98). For furthervalidation we compared the differentiation time course to published microarraydata of TH17 cells and naive T cells51 (Supplementary Fig. 1c). In an additionaldata set, naive T cells were isolated from wild-type and Il23r2/2 mice, and treatedwith IL-6, TGF-b1 and IL-23 and profiled at four different time points (49 h, 54 h,65 h, 72 h). Expression data were pre-processed using the RMA algorithm fol-lowed by quantile normalization52.Detecting differentially expressed genes. Differentially expressed genes (com-paring to the TH0 control) were found using four methods: (1) Fold change.Requiring a twofold change (up or down) during at least two time points. (2)Polynomial fit. We used the EDGE software53,54, designed to identify differentialexpression in time course data, with a threshold of q-value # 0.01. (3) Sigmoidalfit. We used an algorithm similar to EDGE while replacing the polynomials with asigmoid function, which is often more adequate for modelling time course geneexpression data55. We used a threshold of q-value # 0.01. (4) ANOVA. Geneexpression is modelled by: time (using only time points for which we have more

than one replicate) and treatment (‘TGF-b1 1 IL-6’ or ‘TH0’). The model takesinto account each variable independently, as well as their interaction. We reportcases in which the P value assigned with the treatment parameter or the inter-action parameter passed an FDR threshold of 0.01.

Overall, we saw substantial overlap between the methods (average of 82%between any pair of methods). We define the differential expression score of agene as the number of tests that detected it. As differentially expressed genes, wereport cases with differential expression score .2.

For the Il23r2/2 time course (compared to the wild-type T cells) we usedmethods (1)–(3) (above). Here we used a fold change cutoff of 1.5, and reportgenes detected by at least two tests.Clustering. We considered several ways for grouping the differentially expressedgenes, based on their time course expression data: (1) for each time point, wedefined two groups ((a) all the genes that are overexpressed, and (b) all the genesthat are under-expressed relative to TH0 cells (see below)); (2) for each time point,we defined two groups ((a) all the genes that are induced, and (b) all the genes thatare repressed, comparing to the previous time point); (3) k-means clusteringusing only the TH17 polarizing conditions. We used the minimal k, such thatthe within-cluster similarity (average Pearson correlation with the cluster’s cen-troid) was higher than 0.75 for all clusters; and, (4) k-means clustering using aconcatenation of the TH0 and TH17 profiles.

For methods (1) and (2), to decide whether to include a gene, we considered itsoriginal mRNA expression profiles (TH0, TH17) and their approximations assigmoidal functions55 (thus filtering transient fluctuations). We require that thefold change levels (compared to TH0 (method 1) or to the previous time point(method 2)) pass a cutoff defined as the minimum of the following three values:(1) 1.7; (2) mean 1 s.d. of the histogram of fold changes across all time points; or(3) the maximum fold change across all time points. The clusters presented inFig. 1b were obtained with method (4). The groupings from methods (1), (2) and(4) are provided in Supplementary Table 2.Regulatory network inference. We identified potential regulators of TH17 dif-ferentiation by computing overlaps between their putative targets and sets ofdifferentially expressed genes grouped according to methods (1)–(4) above. Weassembled regulator–target associations from several sources: (1) in vivo DNAbinding profiles (typically measured in other cells) of 298 transcriptional regu-lators12–17; (2) transcriptional responses to the knockout of 11 regulatory pro-teins6,43,49,56–60; (3) additional potential interactions obtained by applying theOntogenet algorithm (V. Jojic et al., submitted; regulatory model available at:http://www.immgen.org/ModsRegs/modules.html) to data from the mouseImmGen consortium (http://www.immgen.org; January 2010 release19), whichincludes 484 microarray samples from 159 cell subsets from the innate andadaptive immune system of mice; (4) a statistical analysis of cis-regulatory ele-ment enrichment in promoter regions18; and (5) the transcription factor enrich-ment module of the IPA software (http://www.ingenuity.com/). For everytranscription factor in our database, we computed the statistical significance ofthe overlap between its putative targets and each of the groups defined aboveusing a Fisher’s exact test. We include cases where P , 5 3 1025 and the foldenrichment .1.5.

Each edge in the regulatory network was assigned a time stamp based on theexpression profiles of its respective regulator and target nodes. For the targetnode, we considered the time points at which a gene was either differentiallyexpressed or significantly induced or repressed with respect to the previous timepoint (similarly to grouping methods (1) and (2) above). We defined a regulatornode as ‘absent’ at a given time point if: (i) it was under expressed compared toTH0; or (ii) the expression is low (,20% of the maximum value in time) and thegene was not overexpressed compared to TH0; or, (iii) up to this point in time thegene was not expressed above a minimal expression value of 100. As an additionalconstraint, we estimated protein expression levels using the model from ref. 20and using a sigmoidal fit55 for a continuous representation of the temporalexpression profiles, and the ProtParam software61 for estimating protein half-lives. We require that, in a given time point, the predicted protein level be no lessthan 1.7-fold below the maximum value attained during the time course, and notbe less than 1.7-fold below the TH0 levels. The timing assigned to edges inferredbased on a time-point-specific grouping (grouping methods (1) and (2) above)was limited to that specific time point. For instance, if an edge was inferred basedon enrichment in the set of genes induced at 1 h (grouping method (2)), it will beassigned a ‘1 h’ time stamp. This same edge could then only have additional timestamps if it was revealed by additional tests.Selection of nanostring signature genes. The selection of the 275-gene signature(Supplementary Tables 5 and 6) combined several criteria to reflect as manyaspects of the differentiation program as was possible. We defined the followingrequirements: (1) the signature must include all of the transcription factors thatbelong to a TH17 microarray signature (comparing to other CD41 T cells51, see

ARTICLE RESEARCH

Macmillan Publishers Limited. All rights reserved©2013

Supplementary Methods); that are included as regulators in the network and havea differential expression score .1; or that are strongly differentially expressed(differential expression score 5 4); (2) it must include at least 10 representativesfrom each cluster of genes that have similar expression profiles (using clusteringmethod (4) above); (3) it must contain at least 5 representatives from the pre-dicted targets of each transcription factor in the different networks; (4) it mustinclude a minimal number of representatives from each enriched Gene Ontology(GO) category (computed across all differentially expressed genes); and (5) itmust include a manually assembled list of ,100 genes that are related to thedifferentiation process, including the differentially expressed cytokines, cytokinereceptors and other cell surface molecules. Because these different criteria mightgenerate substantial overlaps, we used a set-cover algorithm to find the smallestsubset of genes that satisfies all of five conditions. We added to this list 18 geneswhose expression showed no change (in time or between treatments) in themicroarray data.

The 86-gene signature (used for the Fluidigm BioMark qPCR assay) is a subsetof the 275-gene signature, selected to include all the key regulators and cytokinesdiscussed. We added to this list 10 control genes (Supplementary Table 5).Selection of perturbation targets. We used an unbiased approach to rank can-didate regulators—transcription factor or chromatin modifier genes—of TH17differentiation. Our ranking was based on the following features: (a) whether thegene encoding the regulator belonged to the TH17 microarray signature (com-paring to other CD41 T cells51, see Supplementary Methods); (b) whether theregulator was predicted to target key TH17 molecules (IL-17, lL-21, IL-23r andROR-ct); (c) whether the regulator was detected based on both perturbation andphysical binding data from the IPA software (http://www.ingenuity.com/); (d)whether the regulator was included in the network using a cutoff of at least 10target genes; (e) whether the gene coding for the regulator was significantlyinduced in the TH17 time course—we only consider cases where the inductionhappened after 4 h to exclude nonspecific hits; (f) whether the gene encoding theregulator was differentially expressed in response to TH17-related perturbationsin previous studies. For this criterion, we assembled a database of transcriptionaleffects in perturbed TH17 cells, including: knockouts of Batf (ref. 56), Rorc (S.Xiao et al., unpublished), Hif1a (ref. 57), Stat3 and Stat5 (refs 43, 62), Tbx21 (A.Awasthi et al., unpublished), Il23r (this study), and Ahr (ref. 59). We also includeddata from the TH17 response to digoxin63 and halofuginone64, as well as informa-tion on direct binding by ROR-ct as inferred from ChIP-seq data (S. Xiao et al.,unpublished). For each regulator, we counted the number of conditions in whichit came up as a significant hit (up/downregulated or bound); for regulators with 2to 3 hits (quantiles 3 to 7 out of 10 bins), we then assigned a score of 1; forregulators with more than 3 hits (quantiles 8–10), we assigned a score of 2 (ascore of 0 is assigned otherwise); and, (g) the differential expression score of thegene in the TH17 time course.

We ordered the regulators lexicographically by the above features according tothe order: (a), (b), (c), (d), (sum of (e) and (f)), (g); that is, first sort according to (a)then break ties according to (b), and so on. We exclude genes that are not over-expressed during at least one time point. As an exception, we retained predictedregulators (features (c) and (d)) that had additional external validation (feature(f)). To validate this ranking, we used a supervised test: we manually annotated 72regulators that were previously associated with TH17 differentiation. All of thefeatures are highly specific for these regulators (P , 1023). Moreover, using asupervised learning method (Naive Bayes), the features provided good predictiveability for the annotated regulators (accuracy of 71%, using fivefold cross valid-ation), and the resulting ranking was highly correlated with our unsupervisedlexicographic ordering (Spearman correlation .0.86).

We adapted this strategy for ranking protein receptors. To this end, we excludedfeature (c) and replaced the remaining ‘protein-level’ features ((b) and (d)) with thefollowing definitions: (b) whether the respective ligand is induced during the TH17time course; and, (d) whether the receptor was included as a target in the networkusing a cutoff of at least 5 targeting transcriptional regulators.Gene knockdown using silicon nanowires. 4 3 4 mm silicon nanowire sub-strates were prepared and coated with 3ml of a 50mM pool of four siGENOMEsiRNAs (Dharmcon) in 96-well tissue culture plates, as previously described10.Briefly, 150,000 naive T cells were seeded on siRNA-laced nanowires in 10ml ofcomplete media and placed in a cell culture incubator (37 uC, 5% CO2) to settle for45 min before full media addition. These samples were left undisturbed for 24 h toallow target transcript knockdown. Afterward, siRNA-transfected T cells wereactivated with anti-CD3/CD28 dynabeads (Invitrogen), according to the manu-facturer’s recommendations, under TH17 polarization conditions (TGF-b1 andIL-6, as above). 10 or 48 h post-activation, culture media was removed from eachwell and samples were gently washed with 100ml of PBS before being lysed in 20mlof buffer TCL (Qiagen) supplemented with 2-mercaptoethanol (1:100 by volume).

After mRNA was collected in Turbocapture plates (Qiagen) and converted tocDNA using Sensiscript RT enzyme (Qiagen), qRT–PCR was used to validate bothknockdown levels and phenotypic changes relative to 8–12 non-targeting siRNAcontrol samples, as previously described65. A 60% reduction in target mRNA wasused as the knockdown threshold. In each knockdown experiment, each individualsiRNA pool was run in quadruplicate; each siRNA was tested in at least threeseparate experiments (Supplementary Fig. 9).mRNA measurements in perturbation assays. We used the nCounter system,presented in full in ref. 66, to measure a custom CodeSet constructed to detect atotal of 293 genes, selected as described above. We also used the FluidigmBioMark HD system to measure a smaller set of 96 genes. Finally, we usedRNA-seq to follow up and validate 12 of the perturbations. Details of the experi-mental and analytical procedures of these analyses are provided in theSupplementary Methods.Profiling Tsc22d3 DNA binding using ChIP-seq. ChIP-seq for Tsc22d3 wasperformed as previously described67 using an antibody from Abcam. The analysisof this data was performed as previously described7 and is detailed in theSupplementary Methods.Estimating statistical significance of monochromatic interactions betweenmodules. The functional network in Fig. 4b consists of two modules: positiveand negative. We compute two indices: (1) within-module index: the percentageof positive edges between members of the same module (that is, downregulationin knockdown/knockout); and (2) between-module index: the percentage ofnegative edges between members of different modules. We shuffled the network1,000 times, while maintaining the nodes’ out degrees (that is, number of out-going edges) and edges’ signs (positive/ negative), and re-computed the twoindices. The reported P values were computed using a t-test.

46. Kim, U. et al. The B-cell-specific transcription coactivator OCA-B/OBF-1/Bob-1 isessential for normal production of immunoglobulin isotypes. Nature 383,542–547 (1996).

47. Wang, V. E., Tantin, D., Chen, J. & Sharp, P. A. B cell development andimmunoglobulin transcription in Oct-1-deficient mice. Proc. Natl Acad. Sci. USA101, 2005–2010 (2004).

48. Bettelli, E. et al. Reciprocal developmental pathways for the generation ofpathogenic effector TH17 and regulatory T cells. Nature 441, 235–238 (2006).

49. Awasthi, A. et al. A dominant function for interleukin 27 in generating interleukin10-producing anti-inflammatory T cells. Nature Immunol. 8, 1380–1389 (2007).

50. Awasthi, A. et al. Cutting edge: IL-23 receptor gfp reporter mice reveal distinctpopulations of IL-17-producing cells. J. Immunol. 182, 5904–5908 (2009).

51. Wei, G. et al. Global mapping of H3K4me3 and H3K27me3 reveals specificity andplasticity in lineage fatedeterminationofdifferentiatingCD41 Tcells. Immunity30,155–167 (2009).

52. Reich, M. et al. GenePattern 2.0. Nature Genet. 38, 500–501 (2006).53. Storey, J., Xiao, W., Leek, J., Tompkins, R. & Davis, R. Significance analysis of time

course microarray experiments. Proc. Natl Acad. Sci. USA 102, 12837–12482(2005).

54. Leek, J. T., Monsen, E., Dabney, A. R. & Storey, J. D. EDGE: extraction and analysis ofdifferential gene expression. Bioinformatics 22, 507–508 (2006).

55. Chechik, G. & Koller, D. Timing of gene expression responses to environmentalchanges. J. Comput. Biol. 16, 279–290 (2009).

56. Schraml, B. U. et al. The AP-1 transcription factor Batf controls TH17differentiation. Nature 460, 405–409 (2009).

57. Shi, L. Z. et al. HIF1a-dependent glycolytic pathway orchestrates a metaboliccheckpoint for the differentiation of TH17 and Treg cells. J. Exp. Med. 208,1367–1376 (2011).

58. Durant, L. et al. Diverse targets of the transcription factor STAT3 contribute to T cellpathogenicity and homeostasis. Immunity 32, 605–615 (2010).

59. Jux, B., Kadow, S. & Esser, C. Langerhans cell maturation and contacthypersensitivity are impaired in aryl hydrocarbon receptor-null mice. J. Immunol.182, 6709–6717 (2009).

60. Xiao, S. et al. Retinoic acid increases Foxp31 regulatory T cells and inhibitsdevelopment of Th17 cells by enhancing TGF-b-driven Smad3 signaling andinhibiting IL-6 and IL-23 receptor expression. J. Immunol. 181, 2277–2284(2008).

61. Wilkins, M. R. et al. Protein identification and analysis tools in the ExPASy server.Methods Mol. Biol. 112, 531–552 (1999).

62. Durant, L. et al. Diverse targets of the transcription factor STAT3 contribute to T cellpathogenicity and homeostasis. Immunity 32, 605–615 (2010).

63. Huh, J. R. et al. Digoxin and its derivatives suppress TH17 cell differentiation byantagonizing RORct activity. Nature 472, 486–490 (2011).

64. Sundrud, M. S. et al. Halofuginone inhibits TH17 cell differentiation by activatingthe amino acid starvation response. Science 324, 1334–1338 (2009).

65. Chevrier, N. et al. Systematic discovery of TLR signaling components delineatesviral-sensing circuits. Cell 147, 853–867 (2011).

66. Geiss, G. K. et al. Direct multiplexed measurement of gene expression with color-coded probe pairs. Nature Biotechnol. 26, 317–325 (2008).

67. Ram, O. et al. Combinatorial patterning of chromatin regulators uncovered bygenome-wide location analysis in human cells. Cell 147, 1628–1639 (2011).

RESEARCH ARTICLE

Macmillan Publishers Limited. All rights reserved©2013